AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Humanline: Online Alignment as Perceptual Loss

Sijia Liu, Niklas Muennighoff, Kawin Ethayarajh

ICLR 2026 · Reinforcement learning › RL from feedback · approach: Reward modeling versus direct optimization

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use humanline variant on offline dataDPO/KTO/GRPO (Llama3-8B-Instruct) · AlpacaEval2 winratewinrate: up to 1.6x higher than offlineImproved
use humanline variant on online dataDPO/KTO/GRPO · AlpacaEval2 winratewinrateNo clear effect
add humanline syncing (ablation)offline+humanline variant · AlpacaEval2 winratewinrateImproved
add humanline clipping on top of syncing (ablation)offline+humanline variant · AlpacaEval2 winratewinrateImproved
humanline GRPO with data sampled 64x less oftenGRPO (Qwen2.5-1.5B-Instruct) · MATH500Pass@1: 0.593 vs 0.593No clear effect
sample data 64x less often without humanline (ablation)GRPO (Qwen2.5-1.5B-Instruct) · MATH500reward / accuracyHurt
sync reference every step (k=1) in math (ablation)humanline GRPO · MATH500training rewardHurt
sync every k in [12, 24] in math (ablation)humanline GRPO · MATH500training rewardImproved
train on lowest quartile of offline data by token log-prob (ablation)offline+humanline DPO · AlpacaEval2 winratewinrateHurt
use humanline sampling instead of humanline clipping (ablation)humanline variants · AlpacaEval2 winratewinrateNo clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy